Efficient deep learning models for oral squamous cell carcinoma classification in histopathological images

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Computational biology

Background:

• Accurate classification of oral squamous cell carcinoma (OSCC) is vital for patient outcomes.
• Manual histopathological examination is the current standard but is subjective and labor-intensive.
• Deep learning offers potential for automated, objective, and efficient image analysis.

Purpose of the Study:

• To evaluate and compare the performance of four deep learning convolutional neural network (CNN) models for binary classification of OSCC.
• To identify the most effective CNN architecture for distinguishing benign from malignant oral histopathological images.

Main Methods:

• A dataset of 10,000 histopathological images of oral lesions was used.
• Four CNN models were investigated: ResNet50, DenseNet201, EfficientNetB0, and ConvNeXt_Tiny.
• Models were trained and tested for binary classification (benign vs. carcinoma).

Main Results:

• EfficientNetB0 demonstrated the highest classification accuracy (97.6%) and ROC-AUC (0.9963).
• ConvNeXt_Tiny achieved 95.92% accuracy, followed by DenseNet201 (86.08%) and ResNet50 (71.52%).
• Modern CNN architectures showed superior performance compared to traditional residual networks.

Conclusions:

• Deep learning, particularly EfficientNetB0, shows high potential for accurate and efficient OSCC classification in digital pathology.
• Integration of advanced CNN models can enhance diagnostic accuracy and reduce observer variability in OSCC detection.
• These findings support the development of AI-assisted diagnostic tools for oral cancer.